Skip to content

AGENTS.md

Latest commit

 

History

History
69 lines (53 loc) · 4.18 KB

AGENTS.md

File metadata and controls

69 lines (53 loc) · 4.18 KB

agent-tools

Agent Inventory

12 agents across 28 skills, following a hub-and-spoke coordination model.

Shared Agents (via Dispatcher Skills)

Agent Dispatcher Skill Consumers Model Tier Tools
code-explorer code-exploration deep-analysis, bug-killer, docs-manager, codebase-analysis, create-spec Sonnet Read, Glob, Grep, Bash
code-architect code-architecture feature-dev, codebase-analysis Sonnet Read, Glob, Grep
researcher research create-spec Sonnet Read, Glob, Grep, Bash

Private Agents (Owned by Workflow Skills)

Agent Owning Skill Model Tier Tools Purpose
code-synthesizer deep-analysis Opus Read, Glob, Grep, Bash Merges explorer findings; deep investigation via git/deps
code-reviewer feature-dev Opus Read, Glob, Grep Confidence-scored code review (threshold: 80)
bug-investigator bug-killer Sonnet Read, Glob, Grep, Bash Evidence-gathering diagnostic agent
docs-writer docs-manager Sonnet Read, Glob, Grep, Bash MkDocs + GitHub-flavored markdown generation
codebase-understanding mr-reviewer Sonnet Read, Glob, Grep Convention, architecture, integration-risk analysis
mr-code-quality mr-reviewer Sonnet Read, Glob, Grep, Bash Language-specific bug/quality review
git-history mr-reviewer Sonnet Read, Glob, Grep, Bash Regression risk via git history analysis
changelog-manager release-python-package Sonnet Bash, Read, Edit, Glob, Grep CHANGELOG.md entry management
task-executor execute-tasks Sonnet Read, Write, Edit, Glob, Grep, Bash 4-phase task execution (Understand, Implement, Verify, Complete)

Coordination Patterns

Hub-and-Spoke — All workflow skills use this topology:

  • Lead assigns work to agents
  • Agents work independently (no cross-agent communication)
  • Results flow back to lead or synthesizer

Model Tiering:

  • Opus — Synthesis, architecture, review (complex cross-cutting reasoning)
  • Sonnet — Exploration, investigation (parallel breadth)

Safety Boundary:

  • No agent has Write/Edit access except task-executor and changelog-manager
  • All other file modifications are performed by the orchestrating lead

Orchestration Topologies

Skill Topology Agent Count
deep-analysis Hub-and-spoke 2-4 explorers + 1 synthesizer
feature-dev Sequential phases, parallel within 2-3 architects + 3 reviewers
bug-killer Track-based branching 2-3 explorers + 1-3 investigators
mr-reviewer Pure parallel batch 3 fixed agents
docs-manager Sequential phases, parallel gen N docs-writers
execute-tasks Wave-based parallel N task-executors (up to max_parallel)

Consumer Applications

Task Manager (apps/task-manager/) — React + Node.js/Express web app that visualizes agent output in real-time.

Agent/Skill Output Files Written Task Manager Consumer
task-executor (execute-tasks) .agents/tasks/{status}/{group}/task-N.json KanbanBoard, TaskDetailPanel, DependencyGraph
execute-tasks orchestrator __live_session__/progress.md WaveProgress (via use-wave-progress hook)
execute-tasks orchestrator __live_session__/execution_context.md ExecutionContextMonitor (via use-execution-context hook)
execute-tasks orchestrator __live_session__/task_log.md SessionTimeline (via use-session-timeline hook)
execute-tasks orchestrator __live_session__/result-*.md ResultPanel (via use-result-file-events hook)
execute-tasks orchestrator __live_session__/session_summary.md SessionHistoryBrowser
create-tasks .agents/tasks/_manifests/{group}.json ProjectSidebar (task group counts)

Interface contract: Agents write JSON/markdown files to .agents/tasks/ and .agents/sessions/. The task-manager watches these directories via Node.js chokidar file watcher (100ms debounce) and broadcasts changes over WebSocket to the React frontend. No direct agent-to-app communication — the filesystem is the sole integration point.